Recall

Skill: Recall

Load the project context scaffold (.claude/context/) at the start of a session, or for a specific task.

Scope: this skill loads the project context scaffold only. Cross-session session memory — the auto-memory memory/MEMORY.md index — is loaded automatically by the harness at session start, so /recall complements it rather than being the whole session-start story.

When to Use

Run /recall at the beginning of a session, especially when:

• Starting work on a project you haven't touched recently
• Picking up a task that touches unfamiliar parts of the codebase
• Onboarding or context-switching between projects

Also useful mid-session: /recall <task description> to load context relevant to a specific task.

Process

Step 1: Check Scaffold Exists

Look for .claude/context/ROUTER.md.

If it doesn't exist: Check for .claude/architecture.md (monolithic fallback).

• If the monolithic file exists, read and summarize it. Suggest running /intel to generate the full scaffold.
• If neither exists, tell the user: "No project context found. Run /intel to generate the context scaffold."

If it exists: Continue to Step 2.

Step 2: Read the Router

Read .claude/context/ROUTER.md to understand what context files are available.

Step 3: Route by Arguments

No arguments → Overview mode:

1. List all context files with:
   • File name and purpose (from ROUTER.md)
   • Last updated date (from each file's header comment)
   • Line count (as a rough size indicator)
   • Status: "Populated" if the file has real content beyond the template, "Template" if still placeholder
2. Flag any files not updated in 30+ days as potentially stale
3. Print a one-line suggestion: "Use /recall <task> to load context for a specific task, or /intel to refresh stale files."

With task description → Targeted mode:

1. Match the task description against the Task Routing table in ROUTER.md
2. Read the matched context files (usually 2-3 files)
3. Print a concise summary of the loaded context:
   • Key architectural facts relevant to the task
   • Relevant conventions to follow
   • Relevant decisions that constrain the approach
   • Any patterns that apply
4. Keep the summary under 30 lines — the point is to prime context, not dump everything

Matching Heuristics

When matching a task description to the routing table:

If the task description doesn't clearly match, default to Onboarding / Context (load everything).

Output Format

Overview mode:

## Project Context Summary

| File | Last Updated | Lines | Status |
|------|-------------|-------|--------|
| architecture.md | 2026-03-15 | 45 | Populated |
| decisions.md | 2026-03-15 | 32 | Populated |
| conventions.md | 2026-03-10 | 28 | Populated |
| patterns/api-endpoints.md | 2026-03-12 | 35 | Populated |

⚠ conventions.md not updated in 20 days — consider running `/intel` to refresh.

Use `/recall <task>` to load context for a specific task.

Targeted mode:

## Context for: [task description]

**Architecture:** [1-2 key facts]
**Conventions:** [relevant rules]
**Decisions:** [relevant constraints]
**Patterns:** [applicable patterns, if any]

Token Discipline

This skill is designed to be lightweight. Do NOT:

• Read every file in full and dump it into the conversation
• Include template placeholder content in summaries
• Read pattern files that aren't relevant to the task
• Summarize more than what's needed for the immediate task

The goal is to prime your context with the minimum needed to work effectively — not to replicate the scaffold in the conversation.